人工神经网络模型基于增强CT影像征象在评估胃癌腹膜微转移中的应用

Artificial neural networks model for prediction of peritoneal micro metastasis of gastric cancer basing on enhanced CT imaging signs

目的:应用人工神经网络(artificial neural network,ANN)模型,通过对增强CT多种征象分析,预测胃癌腹膜微转移情况,建立胃癌腹膜微转移预测模型。方法:收集经手术病理证实的,临床分期为Ⅲ~Ⅳ期的胃癌患者120例,患者术前均接受增强CT检查。根据手术病理是否存在腹膜微转移,分为转移组和未转移组。观察测量治疗前增强CT的基本征象(肿瘤最大径、厚度、浆膜浸润、强化程度、淋巴结转移分站、有无远处转移),间接征象(腹膜周围积液,腹腔脂肪密度增高,腹膜增厚或可疑微结节,网膜浑浊)。首先以单因素分析筛选两组间差异有统计学意义的指标,进一步建立ANN和Logistic回归分析模型判断腹膜是否存在转移。结果 :120例胃癌患者中,手术病理证实共86例存在腹膜转移,34例无腹膜转移。单因素分析显示肿瘤最大径、浆膜浸润、强化程度、淋巴结转移分站、有无淋巴结转移、腹膜周围积液、腹腔脂肪密度增高,网膜浑浊,腹膜增厚或可疑微结节在腹膜转移组及非转移组中差异有统计学意义;Logistic回归分析的总准确率为75.83%(91/120),总敏感度77.91%(67/86),总特异度73.53%(25/34);受试者工作特征曲线(receiver operating characteristic curve,ROC曲线)显示曲线下面积(area under the curve,AUC)Logistic为0.853。ANN模型总准确率88.33%(106/120),总敏感度88.37%(76/86),特异度85.29%(29/34);ROC曲线示AUCANN为0.887。结论:ANN模型可通过对胃癌术前增强CT多种征象分析,为判断术前是否存在腹膜微转移提供一种新的方法,其预测效能高于传统Logistic回归分析。

Objective： To predict the peritoneal micro metastasis of gastric cancer by using artificial neural network（ANN） model basing on enhanced CT imaging signs, and to establish the prediction model of micro metastasis of gastric cancer. Methods： 120 cases of gastric cancer were confirmed by operation and pathology which clinical stage were Ⅲ-Ⅳ, cases of patients with gastric cancer were examined by enhanced CT before operation. According to whether there were peritoneal micro metastasis in operation and pathology, it was divided into metastasis group and non metastasis group. To observe the measurement before treatment enhanced CT signs of（maximal tumor diameter, thickness, serosal invasion, degree of enhancement, substation of lymph node metastases, with or without distant metastasis）, indirect signs（around the peritoneal effusion, higher density of abdominal fat, peritoneal thickening or suspicious micro nodules, omentum turbidity）. First of all, the two groups were screened by single factor analysis, and the ANN and Logistic regression models were estabfished to determine whether the presence of peritoneal metastasis. Results： In 120 cases of gastric cancer, 86 cases were confirmed by operation and pathology, 34 cases had no peritoneal metastasis. Single factor analysis showed that 9 items in the peritoneal metastasis group and non metastasis group has statistically significant difference. There are the maximum tumor diameter, serosal invasion, degree of enhancement, lymph node transfer station, lymph node transfer, peritoneal effusion around, abdominal fat increased density, turbidity of omentum, peritoneal thickening or suspicious micro nodules. Logistic regression analysis shows that the total accuracy 75.83%（91/120）, the total sensitivity of 77.91%（67/86）, specificity is 73.53%（25/34）; ROC curves show that AUCt^c is 0.853. ANN model shows that total accurate rate 88.33%（106/120）, sensitivity 88.37%（76/86）, specificity 85.29% （29/34）;ROC curve shows that AUCANN is 0.887. Conclusions： ANN model can provide a new method to determine the existence of peritoneal micro metastasis before the operation of gastric cancer basing on enhanced CT signs analysis, and the accuracy rate is higher than that of the traditional Logistic regression analysis.